Scalable, Context-Sensitive Psychiatric Assessment With Large Language Models and Brief Audio Diaries

This repository contains the public notebooks and outputs for the paper "Scalable, context-sensitive psychiatric assessment with large language models and brief audio diaries."

Repository Contents

• llm_prediction_per_domain_per_subject_per_day.ipynb
  • Runs domain-level LLM scoring separately for each diary/transcript.
• llm_prediction_reasoning_per_domain.ipynb
  • Selects exemplar high- and low-scoring subjects for each domain.
  • Uses model-based reasoning to generate qualitative summaries.
• Results/llm_prediction_per_domain_per_subject_per_day.csv
  • Released daily diary domain predictions for 108 subjects across 1,337 diary rows.
  • Includes per-domain scores from 6 LLMs plus mean scores across those models.

Domains

The notebooks score five broad DPDS domains:

• Negative Affectivity
• Antagonism
• Detachment
• Disinhibition
• Anankastia

Released Output File

Results/llm_prediction_per_domain_per_subject_per_day.csv contains:

• subject and day identifiers
• Domain predictions from 6 LLMs:
  • claude_sonnet_4
  • gemini_flash
  • gpt_5
  • grok3
  • llama_maverick
  • qwen3_235B
• Mean domain scores across those 6 LLMs:
  • neg_avg
  • ant_avg
  • det_avg
  • dis_avg
  • ank_avg

Running the Notebooks

This is a notebook-first project rather than a packaged Python library.

Typical workflow:

1. Clone the repository.
2. Create a Python environment with Jupyter.
3. Install the notebook dependencies.
4. Create a local Data/ directory with the required restricted inputs.
5. Add your OpenRouter API key in the notebook before running model calls.

Core Python packages used by the notebooks include:

• openai
• tiktoken
• requests
• numpy
• pandas
• jsonschema
• tenacity
• tqdm
• ipython
• jupyter